Unverified50% confidenceFactExact time
RLHF训练中会加入KL散度惩罚项，防止模型偏离SFT阶段学到的基础行为太远
1
Sources
50%
Confidence
Long-term
Relevance
8/21/2026
First Seen
Sources
Related Claims
UnverifiedRLHF和DPO等对齐技术的效果高度依赖预训练基础质量，预训练阶段未习得的能力几乎无法通过后续对齐补救72% similarVerified当代前沿模型的训练方法包括RLHF（基于人类反馈的强化学习）及其后继者RLAIF和过程奖励模型（PRM）71% similarUnverified强化学习训练比监督学习更不稳定，将NVFP4激进量化引入RL训练会进一步放大固有不稳定性70% similarUnverifiedCurrent mainstream LLMs universally adopt a pre-training plus fine-tuning paradigm, first performing unsupervised pre-training on internet-scale corpora, then aligning with human preferences through techniques such as RLHF.69% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/784174API
curl https://kongchang.com/api/v1/knowledge/claims/784174MCP
get_claim(id=784174)